The top 10 agents in Agent Arena process about 93 input tokens for every token they generate, according to our analysis of 26K+ sessions. This is why labs are racing to lower the cost of cached input. The video below shows how caching works in practice, with real sessions from Opus 5.5 by @AnthropicAI and Gemini 4 Argon by @GoogleDeepMind. Dive into the details of the hidden AI price war: article by @DawidGalarowicz below.

Your agent would cost 260% more without the cache. We analyzed +26K Agent Arena sessions from September 12 to October 23, 2026, spanning 30 agents and 13 labs. Caching dramatically changes the economics of a typical task. If every cached token were instead billed at the model's full input price, a task sent to one of Agent Arena's top-10 agents would cost 3.6x as much (260% more). The savings become especially tangible for leading models. A typical Claude Fable 5.1 task costs $7.44 today; without caching, it would cost $27. Even a coffee break can carry a hidden price: if your cache expires while you’re away, the agent may have to resend the entire conversation at full input price. In our Agent Arena sample, a single cache invalidation translated to a 5% higher bill for a typical task. Read more findings from our analysis of the hidden AI price war in the article below.

Arena Conversations: @idorosen, Senior Principal Scientist from @amazon, sat down with @petergostev to discuss practical elements of AI: why unit economics should come first, what a decade in quant trading teaches you about evaluation, and what happens to frontier-lab pricing when open-weight models are only months behind. Hear what Ido says it takes to deploy AI beyond benchmarks: in production, at scale, and on a budget.

Exciting news: Gemini 4 Carbon (High) from @GoogleDeepMind just made a huge debut on Arena, landing #1# in Text Arena, #2# in Agent Arena, and #4# in Code Arena: WebDev. Back at the top, Gemini 4 Carbon delivers frontier intelligence and reshapes the Pareto frontier across three arenas. In Agent Arena, Gemini 4 Carbon ranks #2# with a +16.2% net improvement across +21.9K real-world agentic sessions from our global community, surpassing top models GPT-6 Astra and Fable-5.1. At $1.80 median cost per task, Gemini 4 Carbon (High) costs 72% less than the #1# ranked Claude Opus 5.5 (Max). Its standouts by key signals: - #1# Steerability (+18.6%): can the model course-correct when you push back? - #2# Praise vs Complaint (+37.9%): implicit sentiment in users' reactions - #3# Confirmed Success (+18.7%): explicit “yes, that worked” from users In Code Arena: WebDev, Carbon scored 1768 pts landing #4#: - Just 2 pts below #3# Claude Sonnet 5.5 (xHigh), at 1770 pts - 89 pts above #11# Gemini 4 Argon (High), at 1679 pts By category, it ranks #3# in Gaming and Simulations, and #5# in Brand & Marketing. In Text Arena, Gemini 4 Carbon (High) is the new #1# with 1528 pts, now ahead of Gemini 4 Argon (High) at #2# with 1520 pts. Congrats to the @GoogleDeepMind team on this huge milestone!

Lower cost per task doesn’t always come from lower token prices. On Agent Arena, GPT-6.1 Sol (Max) by @OpenAI costs $1.91 per task: - 85% less than #1# Claude Opus 5.5 (Max) by @AnthropicAI at $12.51 - 87% less than #3# GPT-6 Astra (Max) by @OpenAI, at $14.93 But these savings come from different places. Against Astra, Sol’s cost advantage comes primarily from lower token prices. Against Opus, it comes primarily from using fewer tokens per task. How do we derive this? Agent Arena tracks real costs of users’ tasks. Cost per task is the tokens a model uses times the price of each token, across input, output and cache-read tokens. By comparing pricing and token counts, we can identify what drives the cost reductions.

First impressions of Gemini 4 Carbon’s preview by @GoogleDeepmind are in. @petergostev just put it through a long run of 3D generations, creative prompts, visual explainers and UI tasks, compared against Gemini 4 Argon, Claude Opus 5.5, GPT-6.1 Sol, Muse Wonderful, Kimi and DeepSeek. Watch now to see an early preview of the latest model coming from @GoogleDeepMind. 0:00 Intro: Gemini 4 Carbon 0:15 First test (White House): Carbon vs Argon 1:28 Carbon vs Claude Opus 5.5 2:42 Carbon vs GPT-6.1 Sol 3:06 Carbon vs Muse Wonderful and Kimi 3:48 Why test with 3D generations 4:45 Playable 3D game: Carbon, Argon, Sol, Opus 6:28 Where Carbon falls short: gaps in detail 7:50 Stonehenge 8:30 Golden Gate Bridge 10:41 Golden Gate Bridge: GPT-6.1 Sol comparison 11:53 Carbon's creativity 15:33 Monet paintings transition 18:37 Van Gogh paintings world 21:47 Why creativity matters in a model 23:30 Another 3D prompt: Carbon vs Opus, Sol, Muse, DeepSeek 26:27 London through the ages 29:56 Istanbul 31:44 Harbin ice city 33:20 New test: visual explainers (how lightning works) 35:34 John Snow and the cholera map 38:20 UI work 39:45 Golden Gate Bridge completion animation 40:43 More UI generations 41:55 Ski resort site and ramen ordering 44:20 Verdict

Exciting news: Gemini 4 Carbon (High) from @GoogleDeepMind just made a huge debut on Arena, landing #1# in Text Arena, #2# in Agent Arena, and #4# in Code Arena: WebDev. Back at the top, Gemini 4 Carbon delivers frontier intelligence and reshapes the Pareto frontier across three arenas. In Agent Arena, Gemini 4 Carbon ranks #2# with a +16.2% net improvement across +21.9K real-world agentic sessions from our global community, surpassing top models GPT-6 Astra and Fable-5.1. At $1.80 median cost per task, Gemini 4 Carbon (High) costs 72% less than the #1# ranked Claude Opus 5.5 (Max). Its standouts by key signals: - #1# Steerability (+18.6%): can the model course-correct when you push back? - #2# Praise vs Complaint (+37.9%): implicit sentiment in users' reactions - #3# Confirmed Success (+18.7%): explicit “yes, that worked” from users In Code Arena: WebDev, Carbon scored 1768 pts landing #4#: - Just 2 pts below #3# Claude Sonnet 5.5 (xHigh), at 1770 pts - 89 pts above #11# Gemini 4 Argon (High), at 1679 pts By category, it ranks #3# in Gaming and Simulations, and #5# in Brand & Marketing. In Text Arena, Gemini 4 Carbon (High) is the new #1# with 1528 pts, now ahead of Gemini 4 Argon (High) at #2# with 1520 pts. Congrats to the @GoogleDeepMind team on this huge milestone!

Exciting news: MAI-Image-2.7-Preview from @MicrosoftAI just landed at #3# overall in the Image Edit Arena with 1470 pts! That puts Microsoft AI at #2# among labs. This release is a +45 pt improvement from its previous variant, MAI-Image-2.6 at #7#. MAI-Image-2.7-Preview also improved across every category compared to 2.6: - #3# in Product, Branding & Commercial Design (up from #8#) - #3# in Portraits (up from #9#) - #3# in Photorealistic & Cinematic Imagery (up from #6#) - #4# in Text Rendering (up from #11#) - #4# in Cartoon, Anime & Fantasy (up from #9#) - #5# in 3D Imaging & Modeling (up from #10#) Congrats to the @MicrosoftAI team on this release!

Kimi K3.1 Pro (Max) by @Kimi_Moonshot just landed in the Agent Arena at #11# with +7.6% net improvement across +4K real-world agentic sessions from our global community of users. Kimi K3.1 Pro (Max) would be the new #1# open model in the Agent Arena overtaking DeepSeek-v4.1-Flash! Open weights coming soon, stay tuned. Kimi K3.1 Pro (Max) is a +6 percentage pt improvement over Kimi K3 (Max) at +1.6% (#25# -> #11#). And with a $1.03 median cost per task in Agent Arena, Kimi K3.1 Pro is 13% less than its previous variant. See its placement just off the Pareto frontier in thread. Across Agent Arena categories, Kimi K3.1 Pro (Max) ranks #8# in Work, #9# in Code, and #10# in Chat. By signal, it’s the strongest in Praise vs. Complaint (implicit sentiment in user reactions) with +19% at #8#. Across other signals it landed: - #11# in Steerability (+8.3%) - #15# in Bash Reliability (+8.3%) - #26# in Confirmed Success (+2%) Huge congrats to the @Kimi_Moonshot team on this release!

Kimi K3.1 Pro (Max) by @Kimi_Moonshot just landed in the Agent Arena at #11# with +7.6% net improvement across +4K real-world agentic sessions from our global community of users. Kimi K3.1 Pro (Max) would be the new #1# open model in the Agent Arena overtaking DeepSeek-v4.1-Flash! Open weights coming soon, stay tuned. Kimi K3.1 Pro (Max) is a +6 percentage pt improvement over Kimi K3 (Max) at +1.6% (#25# -> #11#). And with a $1.03 median cost per task in Agent Arena, Kimi K3.1 Pro is 13% less than its previous variant. See its placement just off the Pareto frontier in thread. Across Agent Arena categories, Kimi K3.1 Pro (Max) ranks #8# in Work, #9# in Code, and #10# in Chat. By signal, it’s the strongest in Praise vs. Complaint (implicit sentiment in user reactions) with +19% at #8#. Across other signals it landed: - #11# in Steerability (+8.3%) - #15# in Bash Reliability (+8.3%) - #26# in Confirmed Success (+2%) Huge congrats to the @Kimi_Moonshot team on this release!

Kimi K3.1 is out; it is the #1# OSS model. It places similarly to Opus 5, which was released 3 months ago. Thus, China is lagging the frontier by about 3 months. The Kimi K3 release was #1# on Code Arena, at the time better than all US labs. K3.1 is close, but substantially off from Opus 5.5. On a performance vs cost-per-task basis, the combo of Sonnet 5.5 and DeepSeek V4.1 flash together dominate this model. The race is so tight!!! Congrats to the Moonshot team on a great release. And now, back to work. The frontier never sleeps!

Big news: @Kimi_Moonshot is back in the top-4 labs in the Code Arena: WebDev with the release of Kimi K3.1 Pro! This model just landed #8# in the Code Arena: WebDev with 1692 pts, matching Claude Opus-5 (Max) at 70% lower cost. This is a 9 place jump (+37 points) from Kimi K3 (Max) at #17# with 1655 pts. Kimi K3.1 Pro improved from Kimi K3 across all categories. Highlights: - #7# in Brand & Marketing (up from #9#) - #9# in Simulations (up from #20#) Cost efficiency is another upgrade. Priced at $2/$10 per million input/output tokens, Kimi K3.1 Pro (Max) has a median cost per task of $1.23 (29% less than Kimi K3). See its placement compared to the Pareto frontier in thread. With open weights coming soon, Kimi K3.1 Pro would be the new #1# open model in the Code Arena: WebDev, overtaking DeepSeek-v4.1-Flash. Stay tuned! See @Kimi_Moonshot’s ranking by lab, and 3D generations from Kimi K3.1 Pro below in thread. Huge congrats to the @Kimi_Moonshot team on this release!

Big news: DeepSeek-V4.1-Flash-20261017 (Max) by @deepseek_ai is now the #1# open model in Agent Arena! It just landed at #11# overall with +5.1% net improvement across +7.5K real-world sessions. This performance also reshapes the Agent Arena Pareto frontier! All models above V4.1-Flash 20261017 in performance are proprietary at much higher cost. At $0.20 median cost per task, that’s ~37% of GPT-6.1 Sol’s improvement for about 27% of the cost. By signal, this release achieves: - #8# in Confirmed Success with +11.5% (#1# open) - #11# in Praise vs. Complaint with +5.9% (#1# open) - #12# in Bash Reliability with +5.4% (#1# open) - #13# in Steerability with +2.5% (#1# open) - No issues with Tool Hallucination Find more info on signals at the link in thread on Agent Arena. Congrats to the @deepseek_ai team on this release!

Is it possible for AI to be both creative and effective? The next episode of Arena Conversations is out: @petergostev interviewed Pradeep Natarajan, Senior Principal Scientist at Amazon AGI. They cover: - Model gains from mastering the fundamentals - Evaluations as the differentiator - Harness optimization - Metrics like accuracy and style constraining model creativity Hear about the above and much more.

GPT-6 Luna and GPT-6.1 Sol by @OpenAI now hold 6 of the 11 spots on the Agent Arena Pareto frontier across their reasoning efforts! - GPT-6 Luna dominates below $0.10/task, claiming 3 of the 4 optimal spots with Max, Medium and Low. - GPT-6.1 Sol sets the frontier between $0.22 and $0.65/task, with the same reasoning levels on it too. This shows why test-time scaling matters. Sometimes the best move is to switch the reasoning effort, not the model, to stay at the optimal cost/performance point. With GPT-6 Astra (Max) also on the Pareto frontier, OpenAI models hold 7 of the 11 frontier spots with multiple reasoning levels. Congrats to the @OpenAI team!